How to deploy your first graph in Claude in <60 seconds (bookmark this): The problem with loops is that they run one step at a time. Observe → Act → Verify → Repeat. This approach works fine until your task has multiple independent pieces, and a loop still does them one after another (not efficient at all and too slow). Graphs fix that by running things in parallel instead of in a line of tasks. Here's the exact 4-step process to build one: Step 1. Audit your current loops For each workflow you already run, ask: → How many steps does this take? → Where does it retry most often? → How long does it take and what does it cost? This finds your actual bottlenecks before you change anything. Step 2. Find what can run in parallel For every step, ask one question: does this depend on another step's output? If no, it's a candidate for parallel execution. The most common patterns: research tasks, styling/design, logic/strategy. These usually don't depend on each other the way something like verification does. Step 3. Physically draw the graph Use the Excalidraw MCP, or just hand-draw it. A real example: Task: research 5 competitors and produce a positioning report. As a loop: research competitor 1, then 2, then 3, then write it up, restart if weak. As a graph: → Planner node: defines what to find for each competitor → 5 research nodes: one per competitor, running simultaneously → Synthesizer node: collects all 5 outputs into one report → Review node: checks the report against the original brief Step 4. Send the graph prompt Screenshot your graph and send it to Claude with this prompt attached: Run this as a graph, not a sequence. 1. Plan: define what to research for each competitor. 2. Dispatch 5 subagents IN PARALLEL, one per competitor. Do not run sequentially. 3. Synthesize: collect all 5 outputs into a single report. 4. Review: check the report against the brief. 5. Gate: if review fails, route feedback to the specific node that caused it. Do not re-run the whole graph. Tell me which nodes ran, in what order, and what each returned. Under the hood, this is just subagents with defined roles, dispatched in parallel, with a rule for where failures go. A few things worth knowing: → Start with 3-5 nodes, don't overbuild your first one → Graphs are faster, not always cheaper → /loop is still fine for tasks that don't need parallelization → Be specific about each node's role, vague drawings produce vague graphs(Miles Deutscher)